Incremental Learner for Resource-Constrained Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Resource-constrained devices, such as smartphones and PDAs, face challenges in handling large data and models due to storage and time constraints, requiring fast online algorithms that can update models continuously, but existing machine learning techniques are not designed for these constrained environments.
Innovation Solution
A learner system that generates and updates rules by randomly selecting matching attribute values, assigns weights to rules based on prediction accuracy, and adjusts weights dynamically to improve prediction confidence, allowing for efficient storage and high-speed operation on resource-constrained devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning techniques are used, then model accuracy can be improved, but storage requirements and processing time increase significantly
Solution Approach 1:
The patent segments the learning process into two distinct phases: an offline phase where comprehensive data is processed to build a full model, and an online phase where a simplified incremental learner operates on resource-constrained devices. This segmentation allows the complex model training to occur on powerful servers while the mobile device maintains only a lightweight model that can be continuously updated with minimal storage requirements.
Solution Approach 2:
The patent introduces an intermediary synchronization mechanism that transfers selected model parameters from the offline training process to the online incremental learner. This intermediary process allows the mobile device to benefit from comprehensive offline training while maintaining a compact model structure that fits within device storage constraints, effectively mediating between model accuracy and storage requirements.
2Measurement precision
If comprehensive models are trained, then prediction accuracy improves, but processing speed decreases due to computational complexity
Solution Approach 1:
The patent divides model training into offline batch processing (comprehensive accuracy) and online incremental updates (fast processing). The offline phase trains comprehensive models on servers with unlimited computational resources, while the online phase on mobile devices performs only lightweight incremental updates using the incremental learner, achieving both accuracy and speed requirements.
Solution Approach 2:
The patent implements a dynamic two-phase training architecture where the system adapts between comprehensive offline training and streamlined online learning based on available resources and time constraints. The incremental learner dynamically updates the model with new data points without requiring full model retraining, maintaining processing speed while improving prediction accuracy over time.
3Adaptability or versatility
If models are updated continuously with new data, then adaptability improves, but storage and processing resources are depleted
Solution Approach 1:
The patent extracts only the essential incremental learning operations needed for continuous adaptation, separating them from the computationally intensive full model training. The incremental learner on the mobile device performs only the necessary parameter updates using selected features and new data points, extracting the minimum required processing to maintain adaptability while conserving device resources.
Solution Approach 2:
The patent changes the learning parameters by using different learning rates, feature selection strategies, and update frequencies for the incremental learner compared to traditional batch learning. These parameter adjustments allow the model to adapt continuously to new data with reduced computational overhead, optimizing the balance between adaptability and resource consumption on mobile devices.
Data Source
AI summary
Methods, apparatuses, and articles for registering votes to predict an attribute value for a received instance are described herein. In one embodiment, the registering is performed in a weighted manner based at least on a weight and predicted target values associated with at least one of one or more rules whose antecedent has been met. The meeting of the antecedent is determined based at least on one or more attributes values of one or more other attributes of the received instance. Further, determining whether the predicted target value for which votes are registered correctly predicted the attribute value of the received instance is performed. The associated weight of the rule is adjusted accordingly, for instance, by incrementing the weight if the predicted target value correctly predicted the attribute value of the received instance, and by decrementing the weight if the predicted target value incorrectly predicted the attribute value of the received instance.